Skip to main content
v2026.11,986 entries · CC-BY 4.0

CUSUM Chart: Step-by-Step Implementation, Interpretation, and Analysis Guide

Build and interpret a CUSUM chart for research data in R using the qcc package, with exact decision intervals, reference values, and step-by-step commands.

Written and maintained by CASRAI Editorial Board

Published · Last updated

On this page

A CUSUM chart (cumulative sum control chart) is a time-weighted statistical process control tool designed to detect small, persistent shifts in a process mean or variance by accumulating sequential deviations from a specified target value over time. Unlike standard Shewhart control charts that evaluate each subgroup independently against static three-sigma limits, a CUSUM chart maintains a running memory of historical observations, making it substantially faster and more sensitive at detecting subtle process drifts typically between 0.5 and 2.0 standard errors in laboratory assays, clinical trials, and longitudinal experimental monitoring.

What it is used for in research

In empirical scientific investigations, maintaining experimental stability and measurement fidelity is essential for reproducible research. While large, catastrophic equipment breakdowns or sudden procedural blunders are readily apparent, the most insidious threats to experimental validity are subtle, systematic drifts. These include gradual photodetector decay in spectrophotometers, minor ambient temperature fluctuations in incubator chambers, progressive reagent decay across lot numbers, and cumulative calibration drift in automated liquid-handling robotics. Traditional Shewhart control charts (such as X-bar or individual value charts) are poorly equipped to identify these minor drifts because each plotted point reflects only the immediate sample without context; a persistent 0.75-sigma drift may hover harmlessly within three-sigma Shewhart limits for dozens of consecutive runs without triggering an alert. When selecting an appropriate control chart for clinical and experimental workflows, researchers turn to the cumulative sum control chart specifically to overcome this limitation.

The core statistical mechanism of a CUSUM chart, originally introduced by E.S. Page in 1954, involves plotting cumulative sums of deviations of individual observations or subgroup means from a designated in-control target value (μ0). Mathematically, the raw cumulative sum is defined as:

Sm = Σi=1m (x̄i – μ0)

When an experimental process remains stably centered on the target μ0, positive and negative deviations occur with equal likelihood and cancel each other out, resulting in a cumulative sum that fluctuates randomly in a narrow band around zero. However, when an assignable cause induces even a small, sustained positive or negative mean shift, the individual terms (x̄i – μ0) consistently carry the same sign. The cumulative sum immediately responds by integrating these persistent errors, creating a steep, monotonic slope that rapidly pierces established decision boundaries.

Across research disciplines, CUSUM charts fulfill several critical methodological functions:

  • Analytical Chemistry and Bioassay Validation: Monitoring internal quality control (IQC) reference standards run at the start and end of high-throughput chromatographic or mass spectrometry batches. A CUSUM chart detects micro-drifts in peak area ratios or retention times days before routine analytical runs violate conventional Westgard rules.
  • Clinical Surgery and Procedural Competency: Evaluating procedural learning curves and tracking complication rates. In medical education and clinical audit studies, cumulative sum analysis graphs binary or continuous procedural outcomes to identify when a trainee surgeon achieves procedural competency, or conversely, when institutional complication rates deviate from baseline benchmarks.
  • Environmental Science and Longitudinal Ecology: Detecting baseline regime shifts in continuous sensor networks monitoring atmospheric greenhouse gas fluxes, aquatic dissolved oxygen, and stream pH. CUSUM distinguishes true environmental regime shifts from cyclic seasonal noise.
  • Econometrics and Psychometrics: Serving as a structural break test for continuous time-series models, confirming whether behavioral response latencies or longitudinal panel survey metrics have undergone an unannounced structural shift. While techniques like identifying defect frequencies with Pareto charts isolate common categorical failures, CUSUM charts continuously safeguard continuous process parameters against temporal decay.

The primary metric used to evaluate and compare control chart performance is the Average Run Length (ARL). The in-control run length (ARL0) represents the average number of samples collected before the chart generates a false alarm. Conversely, the out-of-control run length (ARL1) denotes the expected number of samples required to detect a true process shift of magnitude δ = |μ1 – μ0| / σ. For a standard Shewhart X-bar chart with three-sigma limits, ARL0 is approximately 370.4 samples, but detecting a 1.0-sigma shift requires an average of 44.0 samples. A properly calibrated tabular CUSUM chart maintains a comparable or superior in-control ARL0 (between 465 and 930 samples) while detecting that same 1.0-sigma shift in approximately 8 to 10 samples—an efficiency gain of more than fourfold.

How to install or access it

To construct, tune, and evaluate CUSUM charts within a reproducible research pipeline, statistical computing environments provide standardized, peer-reviewed implementations. The most widely adopted open-source tool for statistical process control in research is the qcc (Quality Control Charts) package in R, developed by Luca Scrucca and hosted on the Comprehensive R Archive Network (CRAN). By utilizing open-source statistical software for scientific analysis, research teams eliminate proprietary software licensing barriers while ensuring full scriptability, version control, and reproducible audit trails.

To install qcc directly from CRAN within an active R session (version 3.6 or higher), execute the standard package installer:

# Install the official qcc package from CRAN
install.packages("qcc", dependencies = TRUE)

Once installed, load the package into your environment and verify its build version using the official namespace query:

# Load package and verify installation
library(qcc)
packageVersion("qcc")
# Version 2.7 (or latest CRAN release) will be displayed

In addition to base R package management, the specialized R package spc, maintained by Sven Knoth on CRAN, provides companion numerical routines for evaluating exact zero-state and steady-state Average Run Lengths via Fredholm integral equations and Markov chain representations. Installing spc provides theoretical benchmark calculations alongside experimental charting:

# Install companion package for exact ARL calculations
install.packages("spc")
library(spc)

For researchers working primarily with graphical user interfaces who wish to bypass command-line scripting, JASP (version 0.19 or higher) provides a direct, menu-driven implementation of CUSUM charts. Developed as an open-source initiative supported by the University of Amsterdam, JASP incorporates R’s underlying statistical engines into an accessible desktop interface. JASP can be downloaded directly from its official project repository at https://jasp-stats.org/download/ across Windows, macOS, and Linux platforms. Within JASP, the CUSUM chart is accessible by enabling the Quality Control module from the top-right module manager (+ icon), selecting Control Charts from the top navigation ribbon, and navigating to Time Weighted Charts > CUSUM Chart. While GUI tools offer rapid exploratory scanning, programmatic analysis via R provides superior automation for longitudinal data pipelines.

Step-by-step task

Constructing a rigorous tabular CUSUM chart requires calibrating three interrelated parameters: the process target or center (μ0), the allowable slack or reference value (k), and the decision interval limit (h). In this step-by-step walkthrough, we configure a standardized two-sided tabular CUSUM chart using R’s qcc library, utilizing a standard experimental dataset representing repeated measurements of an analytical calibration standard.

  1. Structure the Experimental Observation Matrix: CUSUM charts can be constructed from individual observations (subgroup size n = 1) or rational subgroups (n ≥ 2). When observations are collected in batches, organize the data into an m × n matrix where rows represent successive time periods and columns represent individual replicate measurements. In R, convert raw vectors with subgroup indicators using qcc.groups():
    # Prepare structured research data
    data(pistonrings, package = "qcc")
    diameter_subgroups <- qcc.groups(pistonrings$diameter, pistonrings$sample)
    dim(diameter_subgroups) # 40 samples, 5 replicates per subgroup
  2. Estimate Baseline Process Parameters: If historical baseline data (Phase I) are available, compute the grand mean and the pooled within-subgroup standard deviation. The parameter center establishes the target μ0, while std.dev defines the process variability σ:
    # Compute baseline Phase I parameters from the first 25 samples
    phase1_data <- diameter_subgroups[1:25, ]
    target_mean <- mean(phase1_data)
    pooled_sd <- stats::sd(phase1_data) # Alternatively use within-subgroup pooled estimators
  3. Set the Reference Value (k) and Shift of Interest: In tabular CUSUM charts, the reference parameter k (allowable slack) determines the magnitude of shift the chart is optimized to detect. For a target shift of δ standard errors, optimal statistical power is achieved by setting k = δ / 2. In qcc::cusum, the argument se.shift specifies δ in units of standard errors of the summary statistics (σ / √n). Setting se.shift = 1.0 automatically establishes k = 0.5 × (σ / √n), tuning the chart to detect a 1-sigma mean shift with maximum efficiency.
  4. Define the Decision Interval Boundary (h): The decision interval decision.interval dictates the alarm threshold expressed in standard error units. The standard industrial and research convention sets decision.interval = 5.0 (or 4.0). A decision interval of h = 5.0 with k = 0.5 yields an in-control Average Run Length (ARL0) of approximately 465 samples for a one-sided chart (or 930 samples two-sided), striking an optimal balance between low false-alarm frequency and rapid shift detection.
  5. Execute the CUSUM Chart Function: Generate the tabular cumulative sum plot by calling the cusum() function. Specify the Phase I dataset, parameters, and plotting flags:
    # Generate baseline CUSUM chart
    cusum_model <- cusum(
      data = phase1_data,
      center = target_mean,
      std.dev = "UWAVE-SD", # Uses unweighted average of subgroup standard deviations
      decision.interval = 5,
      se.shift = 1,
      head.start = 0,
      title = "Two-Sided Tabular CUSUM Chart: Replicate Diameter",
      plot = TRUE
    )
  6. Incorporate Fast Initial Response (FIR) for Startup Monitoring: When restarting equipment, introducing new reagent batches, or commissioning a new experimental assay, the process may already be off-target at startup. The Fast Initial Response (FIR) feature, introduced by Lucas and Crosier (1982), initializes the cumulative sum at head.start = decision.interval / 2 (e.g. 2.5). If the process is in control, the head start quickly decays to zero; if off-target, the chart triggers an out-of-control alarm in roughly half the standard time:
    # Execute CUSUM with Fast Initial Response (FIR)
    cusum_fir <- cusum(
      data = phase1_data,
      center = target_mean,
      decision.interval = 5,
      se.shift = 1,
      head.start = 2.5,
      title = "FIR CUSUM Chart: Head Start at h/2",
      plot = TRUE
    )
  7. Monitor Prospective Phase II Observations: Once Phase I baseline limits are verified, monitor incoming experimental samples by passing the prospective observations to the newdata argument. This holds the original baseline center and control limits constant while projecting new cumulative sums:
    # Monitor prospective Phase II data (samples 26 to 40)
    phase2_data <- diameter_subgroups[26:40, ]
    cusum_phase2 <- cusum(
      data = phase1_data,
      newdata = phase2_data,
      center = target_mean,
      decision.interval = 5,
      se.shift = 1,
      plot = TRUE
    )

Output and how to read it

Interpreting a CUSUM chart requires understanding the mathematical construction of its dual tabular statistics, its control boundaries, and the numerical diagnostics returned by the model summary. Unlike Shewhart charts which plot raw values around a center line, a modern two-sided tabular CUSUM plots two distinct running statistics: the positive (upper) cumulative sum Ci+ and the negative (lower) cumulative sum Ci–.

At each observation or subgroup index i, the standardized tabular cumulative sums are computed recursively according to the following formulas:

Ci+ = max(0, zi – k + Ci-1+)
Ci– = max(0, -zi – k + Ci-1–)

Here, zi = (x̄i – μ0) / (σ / √n) represents the standardized sample score, k is the reference slack value (typically 0.5), and initial values C0+ = C0– = 0 (or h/2 under FIR). The max(0, ...) operator acts as a one-sided statistical rectifier: whenever a sample deviation falls below the reference threshold k, the accumulator resets to zero rather than storing negative values. Consequently, Ci+ only accumulates upward drifts, while Ci– only accumulates downward drifts.

When reviewing the graphical display generated by qcc::cusum, examine the following visual landmarks:

  • Centerline (Zero Line): The baseline origin. In a stable, perfectly centered process, both Ci+ and Ci– frequently return to zero or fluctuate minimally between 0 and 1.5 standard errors.
  • Upper and Lower Decision Boundaries (+h and -h): Horizontal dashed threshold lines positioned at +h (typically +5.0) and -h (-5.0). Any plotted cumulative sum that crosses above +h signals a statistically significant upward process shift. Any plotted cumulative sum crossing below -h signals a significant downward shift. Points crossing these boundaries are rendered in red or highlighted with open circle markers.
  • Trajectory and Change-Point Localization: The slope of the cumulative sum curve reflects the magnitude of the process shift. A steep positive slope indicates a substantial upward shift, whereas a gradual incline indicates a minor, slow drift. Crucially, CUSUM charts allow researchers to retrospectively pinpoint the exact onset of the assignable cause: trace the rising curve backward to the last sample point where Ci was equal to zero. That zero-crossing sample represents the exact time point when the experimental drift began.

To extract numerical diagnostics from the model object, print the summary object using the standard generic method:

# Display comprehensive statistical summary
summary(cusum_model)

The output block provides five critical diagnostic outputs:

  1. Data summary and sample dimensions: Confirms the number of subgroups (m = 25) and subgroup sizes (n = 5).
  2. Target center and standard deviation: Displays the operational center μ0 (e.g. 74.001) and estimated process standard deviation σ (e.g. 0.0098).
  3. Decision parameters: Confirms the calibrated decision.interval (5) and se.shift (1).
  4. Tabular cumulative sums matrix: A matrix displaying the calculated numerical values for Ci+ and Ci– at every successive sample index.
  5. Violating sample index list: An explicit vector listing all sample indices that breached the decision interval boundaries. If no samples violated the threshold, the summary reports that the process is in statistical control.

Historically, researchers utilized a graphical tool known as a V-mask, which placed an angular wedge over the most recent point of an unstandardized cumulative sum plot. While geometrically equivalent under specific parameter transformations, modern international standards (including ISO 7870-4:2021) strongly recommend tabular CUSUM over the V-mask. Tabular CUSUM avoids visual parallax distortion, operates seamlessly in automated software scripts, eliminates manual drafting errors, and provides immediate numerical identification of out-of-control sample points.

The interactive data table below compares the Average Run Length (ARL) performance of tabular CUSUM control charts against standard Shewhart charts across varying process shifts, quantifying the speed advantage of CUSUM for subtle drifts.

Troubleshooting

When implementing CUSUM charts in complex scientific environments, researchers frequently encounter diagnostic discrepancies, software errors, or chart misbehavior. Below are five common operational challenges and their technical resolutions:

  • Excessive False Alarms During Stable Operation: If the CUSUM chart repeatedly breaches decision limits during periods known to be stable, the chart’s in-control Average Run Length (ARL0) is inadequate. This typically occurs when decision.interval is set too low (e.g., h ≤ 3.5), or when underlying experimental observations exhibit positive serial correlation (autocorrelation). In the presence of autocorrelation, consecutive errors compound artificially, driving cumulative sums over the limit. To resolve this, calibrate h to 4.77 or 5.0 for independent data. For autocorrelated sensor or longitudinal data, fit an autoregressive integrated moving average (ARIMA) time-series filter to the data and construct the CUSUM chart on the uncorrelated model residuals.
  • Sluggish Detection of Abrupt, Large Shocks (≥ 3.0 σ): Because CUSUM charts accumulate deviations at an incremental rate determined by k = δ / 2, an abrupt, catastrophic five-sigma spike may require 2 to 3 consecutive samples before the cumulative sum traverses the full decision interval h = 5. To guard against both subtle drifts and sudden large shocks, implement a combined Shewhart-CUSUM monitoring scheme as formalized by Lucas (1982). In this dual framework, standard three-sigma Shewhart upper and lower control limits operate concurrently with the tabular CUSUM chart, ensuring instantaneous alarms for major outliers while preserving cumulative sensitivity for small drifts.
  • Skewed or Non-Normal Measurement Distributions: The theoretical run-length properties of standard CUSUM charts assume normally distributed sample statistics. When monitoring bioassays or cellular concentrations with severe positive skewness, asymmetric false-alarm rates will plague the positive accumulator Ci+. Before charting, verify distributional symmetry by visualizing sample symmetry with box plots in R. When skewness is pronounced, apply an empirical Box-Cox transformation or logarithmic normalization to the raw data prior to calculating subgroup statistics, or deploy a specialized Poisson or negative-binomial CUSUM model designed specifically for discrete count data.
  • R Error ‘differing number of rows’ or Subgroup Matrix Mismatch: In R’s qcc package, passing raw vectors of unequal lengths or unbalanced subgroup data to cusum() without specifying the sizes argument triggers fatal formatting errors. Ensure that input datasets are either structured as balanced numerical matrices where each row represents a complete subgroup, or pass a clean continuous vector paired with an explicit sizes = rep(n, m) argument. Address missing observations (NA) beforehand through documented imputation or sample row exclusion, as qcc does not silently bypass interior missing values.
  • Unreset Accumulators Following Corrective Interventions: A common procedural error is allowing cumulative sums to continue accumulating after an assignable cause has been detected and rectified. Because CUSUM stores historical memory, the accumulator remains elevated near or above the decision boundary even after the process mean returns to target, producing persistent, invalid alarms. Whenever an assignable cause is confirmed and corrected, researchers must explicitly reset the cumulative sum statistics Ci+ and Ci– to zero (or to the FIR head-start value h/2) before resuming monitoring.

Free alternatives

While R and its qcc package represent the benchmark statistical environment for CUSUM modeling in research, investigators have access to several open-source and free alternative platforms. Choosing between them depends on workflow requirements such as graphical interface accessibility, data pipeline integration, and programming language familiarity.

The table below summarizes the leading free and open-source software options supporting CUSUM control charting in academic, clinical, and industrial research environments:

Software Package User Interface CUSUM Capabilities Best Research Use Case License and Access
R (qcc / spc) Command-line / Scripting (RStudio) Tabular CUSUM, FIR head-start, variable subgroup sizes, multivariate CUSUM Automated assay pipelines, reproducible statistical reports, high-throughput batch analysis GPL-2 / GPL-3 (Free, CRAN repository)
JASP (Quality Control) Interactive Graphical (GUI) Tabular and graphical time-weighted CUSUM, interactive parameter sliders Exploratory data analysis, teaching environments, researchers preferring point-and-click GPL-3 (Free, Open-source desktop app)
Python (statsmodels / spc) Command-line / Scripting (Jupyter) CUSUM recursive residual tests for structural breaks, custom SPC chart plotting Machine learning pipelines, econometrics, automated instrumentation backends BSD / MIT (Free, PyPI package index)
jamovi (jqc module) Interactive Graphical (GUI) Standard CUSUM plots powered by underlying R packages, spreadsheet interface Clinical trials and laboratory staff transitioning from spreadsheet software GPL-3 (Free, Open-source desktop app)
GNU Octave Command-line / GUI Scripting Matrix-based CUSUM algorithmic computation, compatible with MATLAB control scripts Engineering laboratories running legacy MATLAB scripts without proprietary licenses GPL-3 (Free, GNU Project)

When comparing R and Stata for research data workflows, R’s community-driven ecosystem provides substantial advantages in specialized statistical process control. In addition to qcc, the specialized R package spc (developed by Sven Knoth) delivers advanced numerical routines for computing exact Average Run Lengths via Markov chain and integral equation approximations. For enterprise Python environments, statsmodels.stats.diagnostic.breaks_cusumolsresid provides econometric CUSUM tests on regression residuals, while general SPC charting can be implemented via lightweight numerical arrays in NumPy and SciPy.

Frequently asked questions

What is the difference between a CUSUM chart and an EWMA chart?

Both CUSUM and EWMA (Exponentially Weighted Moving Average) charts are time-weighted control charts optimized to detect small process shifts (≤ 2.0 σ), but they weight historical data differently. A CUSUM chart assigns equal weight to all past deviations since the last reset, accumulating them linearly until the sum exceeds a decision boundary. In contrast, an EWMA chart applies geometrically decreasing exponential weights to historical observations, giving the greatest influence to the most recent data point. In practice, both charts achieve comparable statistical performance and Average Run Length profiles, but CUSUM offers simpler change-point localization by tracing slopes backward to zero, whereas EWMA provides smoother visual representations of gradual process trends.

When should a researcher choose a CUSUM chart over a Shewhart chart?

A researcher should select a CUSUM chart when the primary analytical objective is detecting small, persistent drifts or step shifts between 0.5 and 2.0 standard deviations from the target mean. Standard Shewhart control charts evaluate each subgroup in isolation and are superior for detecting sudden, large shocks (≥ 3.0 σ). However, Shewhart charts require an average of 44 to 155 samples to detect a 0.5 to 1.0 sigma shift, whereas a calibrated CUSUM chart signals the same shift within 8 to 26 samples. For stability testing in clinical trials, reagent monitoring, and precision calibration, CUSUM provides vastly superior detection speed.

How are the reference value k and decision limit h chosen in practice?

In practical applications, the reference value k (allowable slack) is set to half the anticipated shift magnitude that researchers wish to detect: k = δ / 2, expressed in standard error units (σ / √n). To detect a 1.0-sigma shift in the process mean, set k = 0.5. The decision limit h is typically chosen between 4.0 and 5.0 standard errors. A parameter combination of k = 0.5 and h = 4.0 yields an in-control Average Run Length (ARL0) of approximately 168 samples (one-sided) or 336 samples (two-sided), while h = 5.0 increases ARL0 to approximately 465 samples (one-sided) or 930 samples (two-sided), providing robust protection against false alarms.

What is Fast Initial Response (FIR) in a CUSUM chart?

Fast Initial Response (FIR), introduced by Lucas and Crosier in 1982, is a modification that sets the initial cumulative sum accumulator to a non-zero starting value, typically head.start = h / 2 (e.g. 2.5 when h = 5.0). If a newly initiated or restarted process is in statistical control, the cumulative sum quickly dissipates toward zero with negligible impact on long-term false alarm rates. However, if the process is off-target at startup due to faulty calibration or bad reagents, the head start enables the chart to reach the decision threshold in roughly half the normal time, providing rapid detection of initial startup errors.

Can CUSUM charts be applied to individual observations without subgroups?

Yes, CUSUM charts are exceptionally well-suited for individual observations (subgroup size n = 1), such as discrete batch yields, single analytical patient samples, or daily sensor measurements. When n = 1, the individual values xi are standardized directly using the process mean μ0 and an estimate of the population standard deviation σ. In R’s qcc package, σ for individual observations is typically estimated using the mean moving range of successive observations (MR-bar / d2), ensuring that the cumulative sum statistics are accurately scaled.

How do autocorrelation and serial dependence affect CUSUM charts?

Autocorrelation severely compromises the operational validity of standard CUSUM charts. Positive serial dependence between successive observations causes deviations of the same sign to cluster together, causing cumulative sums to wander spuriously away from zero and triggering frequent false alarms (drastically reducing ARL0). When data exhibit serial correlation, researchers should not plot raw measurements on standard CUSUM charts; instead, fit an autoregressive time-series model (such as AR(1) or ARIMA) to remove temporal dependency, and plot the independent model residuals on the CUSUM chart.

Sources

  • National Institute of Standards and Technology (NIST). NIST/SEMATECH e-Handbook of Statistical Methods: Section 6.3.2.3 Cumulative Sum Control Charts. https://www.itl.nist.gov/div898/handbook/pmc/section3/pmc323.htm (accessed 2026-11-16).
  • National Institute of Standards and Technology (NIST). NIST/SEMATECH e-Handbook of Statistical Methods: Section 6.3.2.3.1 Cusum Average Run Length. https://www.itl.nist.gov/div898/handbook/pmc/section3/pmc3231.htm (accessed 2026-11-16).
  • Comprehensive R Archive Network (CRAN). Package ‘qcc’: Quality Control Charts (Version 2.7 by Luca Scrucca). https://cran.r-project.org/package=qcc (accessed 2026-11-16).
  • Scrucca, L. (2004). qcc: an R package for quality control charting and statistical process control. R News, 4(1), 11-17. https://cran.r-project.org/doc/Rnews/Rnews_2004-1.pdf (accessed 2026-11-16).
  • Page, E.S. (1954). Continuous Inspection Schemes. Biometrika, 41(1/2), 100-115. https://doi.org/10.1093/biomet/41.1-2.100 (accessed 2026-11-16).
  • International Organization for Standardization (ISO). ISO 7870-4:2021 Control charts — Part 4: Cumulative sum charts. https://www.iso.org/standard/76503.html (accessed 2026-11-16).
  • JASP Team. JASP Quality Control Module: Control Charts and Time Weighted Charts (Version 0.19.1). https://jasp-stats.org/ (accessed 2026-11-16).

Data: Average Run Length (ARL) Comparison of CUSUM and Shewhart Control Charts

Comparative compilation of Average Run Length (ARL) performance metrics benchmarking two-sided Tabular CUSUM control charts against standard 3-sigma Shewhart X-bar charts across standardized mean shifts from 0.00 to 4.00 standard errors, compiled from the NIST/SEMATECH e-Handbook of Statistical Methods and statistical process control literature.

Period covered: 1954/2026. Coverage: Global.

Average Run Length (ARL) Comparison of CUSUM and Shewhart Control Charts
Mean Shift (Delta) (sigma/sqrt(n))Shewhart X-bar ARLCUSUM ARL (h=4, k=0.5)CUSUM ARL (h=5, k=0.5)CUSUM vs Shewhart Efficiency Ratio
0.00 (In-Control)3713369300.91
0.25281.1474.21403.79
0.50155.2226.6305.84
0.7581.2213.3176.11
1.00448.3810.45.25
1.5014.974.755.753.15
2.006.33.344.011.89
2.503.242.623.111.24
3.0022.192.570.91
4.001.191.712.010.7

Sources

Compiled by CASRAI from the sources above. Free to reuse under CC BY 4.0 with attribution.

Follow CASRAI

Research-administration guidance, standards updates and independent tool reviews.

Ask CASRAI · free to try

Ask about CUSUM Chart: Step-by-Step Implementation, Interpretation, and Analysis Guide

2 free questions a day, no account.

An AI assistant for research: funder rules, compliance, methods and publishing. It links the sources behind its answers and labels answers that come from a web search.

Answers draw on CASRAI's guides and dictionary plus the federal and funder documents we index: Federal Register, Grants.gov, Regulations.gov and UKRI.

Works on this site and inside Claude, Cursor and the AI tools you already use.

Everything CASRAI publishes — this page, the dictionary, the guides and the news — stays open to read, with no account and no card.

Referenced across the research world

University of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logoUniversity of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logo
  • University of Cambridge logo
  • Columbia University logo
  • Crossref logo
  • University of Edinburgh logo
  • Harvard University logo
  • University of Oxford logo
  • Princeton University logo
  • Stanford School of Medicine logo
  • University College London logo
  • ORCID logo

View CASRAI adoption →